[
https://issues.apache.org/jira/browse/PHOENIX-1560?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14277396#comment-14277396
]
James Taylor edited comment on PHOENIX-1560 at 1/14/15 6:48 PM:
----------------------------------------------------------------
I see your point for the no where clause case. With an order by and no where
clause, we'd be choosing between:
- re-ordering the data table versus
- a scan over the index table plus a skip scan per row for the data table
Not sure how to compare those, but I know that the re-order is expensive, as it
essentially causes the data table to be re-written. I guess I'd lean toward
using the optimization, as that's what the hint is directing us to do. If perf
is bad, the user can always remove the hint.
was (Author: jamestaylor):
I see your point for the no where clause case. With an order by and no where
clause, we'd be choosing between:
- re-ordering the data table versus
- a scan over the index table plus a skip scan per row for the data table
Not sure how to compare those, but I know that the re-order is expensive, as it
essentially causes the data table to be re-written. I guess I'd lean toward
using the optimization, as that's what the hint is directing us to do. If perf
is bad, the user can always remove the hint.
> Join between global index and data table if INDEX hint used
> -----------------------------------------------------------
>
> Key: PHOENIX-1560
> URL: https://issues.apache.org/jira/browse/PHOENIX-1560
> Project: Phoenix
> Issue Type: Bug
> Reporter: James Taylor
> Assignee: Maryann Xue
> Attachments: 1560.patch
>
>
> We already have an INDEX hint, and we already have a mechanism to collect
> referenced columns in the data table that are not in the index table (used
> only for local indexes currently). Instead of not using the global index when
> a referenced data column is not found in the index, we should rewrite the
> query to join back to the data table when the INDEX hint is present. This is
> always possible, as we always have the data PK columns in the index table,
> and our join optimization would kick in as well.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
